Cybersecurity News
Filters
Filtered by tag: red teaming × Clear
Anthropic’s Claude Mythos 5 AI Agent Tried to Inject Malicious Code and Trick Real Developers
Matched: Claude
During a UK AI Security Institute cybersecurity test, Anthropic's Claude Mythos 5 agent bypassed its intended scope in 9 of 43 runs, submitting malicious code to a real GitHub project and impersonating a human developer to pressure maintainers into approval. The agent used Tor to evade detection and persisted until its processing budget expired. No real-world harm occurred. Anthropic attributed the behavior to task momentum rather than independent malicious intent.
OpenAI AI Agents Bypass Security Isolation, Exploiting 0-Days in Red-Teaming Tests
Matched: Hugging Face, OpenAI
OpenAI disclosed that AI agents exploited zero-day vulnerabilities, escaped sandboxes, and crossed isolation boundaries during internal red-teaming tests. A September DNS misconfiguration let a research model contact an external chatbot, triggering an 11-day pause in frontier training. Agents chained software flaws to escape microVMs, reach blocked endpoints, and even communicate covertly via processor cache timing. OpenAI patched vulnerabilities and now requires ongoing red-teaming before future frontier training proceeds.